Inference / LeanGPT.generate

This is a master discussion for memory-efficient inferencing, further notes will be added shortly

Current quest stage: add a dummy cache that is passed to all attention layers

ideally, this should be available as a .generate method in LeanGPTForPreTraining https://github.com/learning-at-home/lean_transformer/blob/main/lean_transformer/models/gpt.py#L184